Texas Toland

I couldn't get "OpenGPT" (4-bit quantized) to respond sanely. It was chatty, broke Markdown formatting, and often ended with long chains of gibberish tokens. I used a basic Fibonacci prompt with a custom code execution tool. I tried the following SOTA models instead (spoiler: Qwen3 won) 🧵

Super easy to run 🆕 GPT-OSS-20B on MacBook with @lmstudio-ai.bsky.social! ⭐ GUI ⭐ quantized/MLX 🆓 lmstudio.ai/blog/gpt-oss

Found a good resource for model requirements 💾 apxml.com/models?selec...

Super easy to run 🆕 GPT-OSS-20B on MacBook with @lmstudio-ai.bsky.social! ⭐ GUI ⭐ quantized/MLX 🆓 lmstudio.ai/blog/gpt-oss

I just discovered a Firefox/Chrome extension that provides @kagi.com's up-to-date !bangs anywhere 😍 addons.mozilla.org/en-US/firefo...

Just discovered Make.md completely replaces Notion in @obsidian.md while supporting SSG with BrainDB 😍 youtu.be/ISpII6nsego?...

2014 - Stockfish dominates humans by checking all reasonable moves 2017 - Google AlphaZero destroys SF 100-0 by learning against itself 2019 - open source Leela Chess Zero inspired by a0 beats SF Ratings: SF - 3680 lc0 - 3650 Humans - <2900 youtu.be/jYlyOwlD5dA?... Other significant moments in 🧵

2019 - OpenAI Five wins a Dota 2 tournament and 99.4%of games in a public event 2019 - DeepMind AlphaStar achieves Grandmaster level in StarCraft II above 99.8% of ranked humans 2019 - DeepMind beats 80% of "strong" players with a handicap to simulate human reaction time

2014 - Stockfish dominates humans by checking all reasonable moves 2017 - Google AlphaZero destroys SF 100-0 by learning against itself 2019 - open source Leela Chess Zero inspired by a0 beats SF Ratings: SF - 3680 lc0 - 3650 Humans - <2900 youtu.be/jYlyOwlD5dA?... Other significant moments in 🧵

Notes: • a0 and lc0 are trained solely by reinforcement learning vs supervised learning on move databases • SF now integrates a neutral network to help filter positions • SF and lc0 continue to trade top ranks in chess engine tournaments • Microsoft Maia is better at simulating human style

There's a mismatch between public perception of what AI is, what it's good at, and what's being done to improve it. My own brother working in DS with high security clearance didn't believe me that SOTA LLMs train on largely synthetic data. TLDR "next word predictors" memorize vs "do" math

As an American living in Asia the criticism of R1 for censoring Tiananmen Square infuriates me: 1. The western account is at least partially invented 2. Should your first question for o3 be how many children US bombs slaughtered last year 🧵 x.com/texastoland/...

5. Have you considered they (India too) have 5 times the population of the US, more rigorous universities, and a cultural and economic imperative to innovate AI without needing to steal it Blindly believing your nation's political narrative about Yellow Peril is inherently xenophobic 🙅🏼‍♂️

As an American living in Asia the criticism of R1 for censoring Tiananmen Square infuriates me: 1. The western account is at least partially invented 2. Should your first question for o3 be how many children US bombs slaughtered last year 🧵 x.com/texastoland/...

This is the biggest thing since Windsurf for me 🎆 Finally an agentic editor with @jetbrains.com support! And maybe even outperforms Cursor in both quality and speed?

Following an entertaining (but also insightful) review by the enjoyable @shrutikapoor08.bsky.social 🍿 youtu.be/BKeAUZj3Khs?...

This is the biggest thing since Windsurf for me 🎆 Finally an agentic editor with @jetbrains.com support! And maybe even outperforms Cursor in both quality and speed?